Hespas Documentation
Hespas is a distributed ML performance estimation tool that takes a StableHLO workload representation of an ML model and produces a time estimate for that workload. It splits the StableHLO programs into compute and communication operators. Compute performance estimation is done using analytical, simulation, or profiling-based estimators. It then outputs a Chakra execution trace annotated with measured timings that is used as input to ASTRA-sim for network simulation.
Quick Start
git clone https://github.com/imec-int/hespas.git
cd hespas
pip install .
Alternatively, you can use uv:
uv sync
Generate a Chakra trace for a sample workload:
hespas_chakra_gen tests/fixtures/configs/config_roofline_a100.json